/
base_module.py
410 lines (370 loc) · 24 KB
/
base_module.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
import os
import argparse
import torch
from torch.utils.data import DataLoader
import pytorch_lightning as pl
import wandb
import sys
import numpy as np
sys.path.append('..')
from datasets.Pendulum import PendulumDataset
from datasets.LotkaVolterra import LotkaVolterraDataset
from datasets.NBody import NBodyDataset
from datasets.PixelPendulum import PixelPendulumDataset
from utilities.losses import *
class BaseDynamicsModule(pl.LightningModule):
def __init__(self, *args, **kwargs):
super().__init__()
self.save_hyperparameters()
if self.hparams.rec_loss_type == 'MSE':
self.reconstruction_loss = mse_loss
elif self.hparams.rec_loss_type == 'L1':
self.reconstruction_loss = l1_loss
elif self.hparams.rec_loss_type == 'BCE_LOGITS':
self.reconstruction_loss = bce_with_logits_loss
elif self.hparams.rec_loss_type == 'CNN_MSE':
self.reconstruction_loss = cnn_vae_mse_loss
else:
raise Exception(f'Wrong loss type {self.hparams.rec_loss_type}')
self.val_rec_loss_sizes = [1, 5, 10, 20]
self.test_rec_loss_sizes = [1, 5, 10, 20, 50, 100, 200]
if self.hparams.model_output_size not in self.val_rec_loss_sizes:
self.val_rec_loss_sizes.append(self.hparams.model_output_size)
self.train_ind = None
self.coord_dim = self.get_coords_dim()
self.num_factors = self.get_factors_dim()
def get_coords_dim(self):
if 'pixel_pendulum' in self.hparams.dataset:
coord_dim = None
elif 'pendulum' in self.hparams.dataset:
coord_dim = 2
elif 'lv' in self.hparams.dataset:
coord_dim = 2
elif '3body' in self.hparams.dataset:
coord_dim = 6
return coord_dim
def get_factors_dim(self):
if 'pixel' in self.hparams.dataset:
factors_dim = 1
elif 'pendulum' in self.hparams.dataset:
factors_dim = 1
elif 'lv' in self.hparams.dataset:
factors_dim = 4
elif '3body' in self.hparams.dataset:
factors_dim = 4
return factors_dim
def prepare_data(self, dataset_dir=os.environ['DATA_DIR_PHYSICS']):
print('Preparing data')
self.datasets = {}
dt = self.hparams.dataset_dt
if self.hparams.dataset=='pendulum_var_length':
train_name = f'pendulum_n_10000_steps_2000_dt_{dt}_len_1.0-1.5_angle_10-170_g_9.81.hd5'
test_name1 = f'pendulum_n_1000_steps_2000_dt_{dt}_len_0.5-1.0_angle_10-170_g_9.81.hd5'
test_name2 = f'pendulum_n_1000_steps_2000_dt_{dt}_len_1.5-2.0_angle_10-170_g_9.81.hd5'
train_path=os.path.join(dataset_dir, train_name)
test_path1=os.path.join(dataset_dir, test_name1)
test_path2=os.path.join(dataset_dir, test_name2)
self.datasets['train'] = PendulumDataset(train_path, self.hparams.coordinates, self.hparams.noise_std, indexes=list(range(0,9000)))
self.datasets['val'] = PendulumDataset(train_path, self.hparams.coordinates, self.hparams.noise_std, indexes=list(range(9000,10000)))
self.datasets['test'] = [self.datasets['val'],
PendulumDataset(test_path1, self.hparams.coordinates, self.hparams.noise_std),
PendulumDataset(test_path2, self.hparams.coordinates, self.hparams.noise_std)]
if self.hparams.dataset=='pendulum-2':
train_name = f'pendulum_n_10000_steps_2000_dt_{dt}_len_1.0-1.5_angle_10-170_g_9.81.hd5'
test_name1 = f'pendulum_n_1000_steps_2000_dt_{dt}_len_0.90-1.00_angle_10-170_g_9.81.hd5'
test_name2 = f'pendulum_n_1000_steps_2000_dt_{dt}_len_1.50-1.60_angle_10-170_g_9.81.hd5'
train_path=os.path.join(dataset_dir, train_name)
test_path1=os.path.join(dataset_dir, test_name1)
test_path2=os.path.join(dataset_dir, test_name2)
rand_ind = np.random.uniform(0.0,1.0,10000)
if self.train_ind is None:
self.train_ind = np.where(rand_ind<=0.80)[0]
self.val_ind = np.where((rand_ind>0.80) & (rand_ind<=0.90))[0]
self.test_ind = np.where(rand_ind>=0.90)[0]
train_set = PendulumDataset(train_path, self.hparams.coordinates, self.hparams.noise_std, indexes=self.train_ind)
val_set = PendulumDataset(train_path, self.hparams.coordinates, self.hparams.noise_std, indexes=self.val_ind)
in_dist_test_set = PendulumDataset(train_path, self.hparams.coordinates, self.hparams.noise_std, indexes=self.test_ind)
self.datasets['train'] = train_set
self.datasets['val'] = val_set
self.datasets['test'] = [train_set,
val_set,
in_dist_test_set,
PendulumDataset(test_path1, self.hparams.coordinates, self.hparams.noise_std),
PendulumDataset(test_path2, self.hparams.coordinates, self.hparams.noise_std)]
elif self.hparams.dataset=='lv-1':
train_name = f'lv_n_10000_steps_2000_dt_{dt}_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05.hd5'
val_name = f'lv_n_1296_steps_2000_dt_{dt}_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05_away_0.00_0.01.hd5'
test_name1 = f'lv_n_1296_steps_2000_dt_{dt}_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05_away_0.01_0.03.hd5'
test_name2 = f'lv_n_1296_steps_2000_dt_{dt}_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05_away_0.03_0.05.hd5'
train_path=os.path.join(dataset_dir, train_name)
val_path=os.path.join(dataset_dir, val_name)
test_path1=os.path.join(dataset_dir, test_name1)
test_path2=os.path.join(dataset_dir, test_name2)
self.datasets['train'] = LotkaVolterraDataset(train_path, self.hparams.noise_std)
self.datasets['val'] = LotkaVolterraDataset(val_path, self.hparams.noise_std)
self.datasets['test'] = [self.datasets['val'],
LotkaVolterraDataset(test_path1, self.hparams.noise_std),
LotkaVolterraDataset(test_path2, self.hparams.noise_std)]
elif self.hparams.dataset=='lv-2':
train_name = f'lv-r_n_10000_steps_1000_dt_0.01_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.10.hd5'
test_name1 = f'lv-r_n_1000_steps_1000_dt_0.01_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.10_away_0.00_0.02.hd5'
test_name2 = f'lv-r_n_1000_steps_1000_dt_0.01_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.10_away_0.02_0.04.hd5'
train_path=os.path.join(dataset_dir, train_name)
test_path1=os.path.join(dataset_dir, test_name1)
test_path2=os.path.join(dataset_dir, test_name2)
rand_ind = np.random.uniform(0.0,1.0,10000)
if self.train_ind is None:
self.train_ind = np.where(rand_ind<=0.80)[0]
self.val_ind = np.where((rand_ind>0.80) & (rand_ind<=0.90))[0]
self.test_ind = np.where(rand_ind>=0.90)[0]
train_set = LotkaVolterraDataset(train_path, self.hparams.noise_std, indexes=self.train_ind)
val_set = LotkaVolterraDataset(train_path, self.hparams.noise_std, indexes=self.val_ind)
in_dist_test_set = LotkaVolterraDataset(train_path, self.hparams.noise_std, indexes=self.test_ind)
self.datasets['train'] = train_set
self.datasets['val'] = val_set
self.datasets['test'] = [train_set,
val_set,
in_dist_test_set,
LotkaVolterraDataset(test_path1, self.hparams.noise_std),
LotkaVolterraDataset(test_path2, self.hparams.noise_std)
]
elif self.hparams.dataset=='lv-3':
train_name = f'lv-r_n_10000_steps_1000_dt_0.01_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05.hd5'
test_name1 = f'lv-r_n_1000_steps_1000_dt_0.01_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05_away_0.00_0.01.hd5'
test_name2 = f'lv-r_n_1000_steps_1000_dt_0.01_X0_5.0_3.0_means_2.0_1.0_4.0_1.0_delta_0.05_away_0.01_0.02.hd5'
train_path=os.path.join(dataset_dir, train_name)
test_path1=os.path.join(dataset_dir, test_name1)
test_path2=os.path.join(dataset_dir, test_name2)
rand_ind = np.random.uniform(0.0,1.0,10000)
if self.train_ind is None:
self.train_ind = np.where(rand_ind<=0.80)[0]
self.val_ind = np.where((rand_ind>0.80) & (rand_ind<=0.90))[0]
self.test_ind = np.where(rand_ind>=0.90)[0]
train_set = LotkaVolterraDataset(train_path, self.hparams.noise_std, indexes=self.train_ind)
val_set = LotkaVolterraDataset(train_path, self.hparams.noise_std, indexes=self.val_ind)
in_dist_test_set = LotkaVolterraDataset(train_path, self.hparams.noise_std, indexes=self.test_ind)
self.datasets['train'] = train_set
self.datasets['val'] = val_set
self.datasets['test'] = [train_set,
val_set,
in_dist_test_set,
LotkaVolterraDataset(test_path1, self.hparams.noise_std),
LotkaVolterraDataset(test_path2, self.hparams.noise_std)
]
elif self.hparams.dataset=='3body-1':
train_name = f'3body_n_10000_steps_1000_dt_{dt}_params_K1_1.00_K2_0.80-1.20_m1_0.80-1.20_m2_0.80-1.20_m3_0.80-1.20.hd5'
val_name = f'3body_n_1296_steps_1000_dt_{dt}_params_K1_1.00_K2_0.80-1.20_m1_0.80-1.20_m2_0.80-1.20_m3_0.80-1.20_away_0.00-0.02.hd5'
test_name = f'3body_n_1296_steps_1000_dt_{dt}_params_K1_1.00_K2_0.80-1.20_m1_0.80-1.20_m2_0.80-1.20_m3_0.80-1.20_away_0.02-0.10.hd5'
train_path=os.path.join(dataset_dir, train_name)
val_path=os.path.join(dataset_dir, val_name)
test_path=os.path.join(dataset_dir, test_name)
self.datasets['train'] = NBodyDataset(train_path, self.hparams.noise_std)
self.datasets['val'] = NBodyDataset(val_path, self.hparams.noise_std)
self.datasets['test'] = [self.datasets['val'],
NBodyDataset(test_path, self.hparams.noise_std)]
elif self.hparams.dataset=='3body-2':
train_name = f'3body_n_10000_steps_1000_dt_0.01_params_K1_1.00_K2_1.00-1.10_m1_1.00-1.10_m2_1.00-1.10_m3_1.00-1.10.hd5'
val_name = f'3body_n_1296_steps_1000_dt_0.01_params_K1_1.00_K2_1.00-1.10_m1_1.00-1.10_m2_1.00-1.10_m3_1.00-1.10_away_0.00-0.01.hd5'
test_name1 = f'3body_n_1296_steps_1000_dt_0.01_params_K1_1.00_K2_1.00-1.10_m1_1.00-1.10_m2_1.00-1.10_m3_1.00-1.10_away_0.01-0.05.hd5'
test_name2= f'3body_n_1296_steps_1000_dt_0.01_params_K1_1.00_K2_1.00-1.10_m1_1.00-1.10_m2_1.00-1.10_m3_1.00-1.10_away_0.05-0.10.hd5'
train_path=os.path.join(dataset_dir, train_name)
val_path=os.path.join(dataset_dir, val_name)
test_path1=os.path.join(dataset_dir, test_name1)
test_path2=os.path.join(dataset_dir, test_name2)
self.datasets['train'] = NBodyDataset(train_path, self.hparams.noise_std)
self.datasets['val'] = NBodyDataset(val_path, self.hparams.noise_std)
self.datasets['test'] = [self.datasets['val'],
NBodyDataset(test_path1, self.hparams.noise_std),
NBodyDataset(test_path2, self.hparams.noise_std)]
elif self.hparams.dataset=='3body-3':
train_name = f'3body_n_10000_steps_1000_dt_0.01_params_K2_1.00_m1_1.00_m2_1.00_m3_1.00_delta_0.05.hd5'
test_name1 = f'3body_n_1000_steps_1000_dt_0.01_params_K2_1.00_m1_1.00_m2_1.00_m3_1.00_delta_0.05_away_0.00_0.01.hd5'
test_name2 = f'3body_n_1000_steps_1000_dt_0.01_params_K2_1.00_m1_1.00_m2_1.00_m3_1.00_delta_0.05_away_0.01_0.02.hd5'
train_path = os.path.join(dataset_dir, train_name)
test_path1 = os.path.join(dataset_dir, test_name1)
test_path2 = os.path.join(dataset_dir, test_name2)
rand_ind = np.random.uniform(0.0,1.0,10000)
if self.train_ind is None:
self.train_ind = np.where(rand_ind<=0.80)[0]
self.val_ind = np.where((rand_ind>0.80) & (rand_ind<=0.90))[0]
self.test_ind = np.where(rand_ind>=0.90)[0]
train_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.train_ind)
val_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.val_ind)
in_dist_test_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.test_ind)
self.datasets['train'] = train_set
self.datasets['val'] = val_set
self.datasets['test'] = [train_set,
val_set,
in_dist_test_set,
NBodyDataset(test_path1, self.hparams.noise_std),
NBodyDataset(test_path2, self.hparams.noise_std)]
elif self.hparams.dataset=='3body-4':
train_name = f'3body_n_10000_steps_1000_dt_0.01_params_K2_2.00_m1_2.00_m2_2.00_m3_2.00_delta_0.05.hd5'
test_name1 = f'3body_n_1000_steps_1000_dt_0.01_params_K2_2.00_m1_2.00_m2_2.00_m3_2.00_delta_0.05_away_0.00_0.01.hd5'
test_name2 = f'3body_n_1000_steps_1000_dt_0.01_params_K2_2.00_m1_2.00_m2_2.00_m3_2.00_delta_0.05_away_0.01_0.02.hd5'
train_path = os.path.join(dataset_dir, train_name)
test_path1 = os.path.join(dataset_dir, test_name1)
test_path2 = os.path.join(dataset_dir, test_name2)
rand_ind = np.random.uniform(0.0,1.0,10000)
if self.train_ind is None:
self.train_ind = np.where(rand_ind<=0.80)[0]
self.val_ind = np.where((rand_ind>0.80) & (rand_ind<=0.90))[0]
self.test_ind = np.where(rand_ind>=0.90)[0]
train_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.train_ind)
val_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.val_ind)
in_dist_test_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.test_ind)
self.datasets['train'] = train_set
self.datasets['val'] = val_set
self.datasets['test'] = [train_set,
val_set,
in_dist_test_set,
NBodyDataset(test_path1, self.hparams.noise_std),
NBodyDataset(test_path2, self.hparams.noise_std)]
elif self.hparams.dataset=='3body-5':
train_name = f'3body_n_10000_steps_1000_dt_0.01_params_K2_2.00_m1_2.00_m2_2.00_m3_2.00_delta_0.10.hd5'
test_name1 = f'3body_n_1000_steps_1000_dt_0.01_params_K2_2.00_m1_2.00_m2_2.00_m3_2.00_delta_0.10_away_0.00_0.02.hd5'
test_name2 = f'3body_n_1000_steps_1000_dt_0.01_params_K2_2.00_m1_2.00_m2_2.00_m3_2.00_delta_0.10_away_0.02_0.05.hd5'
train_path = os.path.join(dataset_dir, train_name)
test_path1 = os.path.join(dataset_dir, test_name1)
test_path2 = os.path.join(dataset_dir, test_name2)
rand_ind = np.random.uniform(0.0,1.0,10000)
if self.train_ind is None:
self.train_ind = np.where(rand_ind<=0.80)[0]
self.val_ind = np.where((rand_ind>0.80) & (rand_ind<=0.90))[0]
self.test_ind = np.where(rand_ind>=0.90)[0]
train_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.train_ind)
val_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.val_ind)
in_dist_test_set = NBodyDataset(train_path, self.hparams.noise_std, indexes=self.test_ind)
self.datasets['train'] = train_set
self.datasets['val'] = val_set
self.datasets['test'] = [train_set,
val_set,
in_dist_test_set,
NBodyDataset(test_path1, self.hparams.noise_std),
NBodyDataset(test_path2, self.hparams.noise_std)]
elif self.hparams.dataset=='pixel_pendulum-1':
train_name = f'pixel_pendulum_n_1296_steps_100_dt_{dt:.2f}_angle_30-170_vel_-2.00-2.00_len_1.20-1.40_g_8.00-12.00.hd5'
val_name = f'pixel_pendulum_n_81_steps_100_dt_{dt:.2f}_angle_30-170_vel_-2.00-2.00_len_1.40-1.45_g_12.00-12.50.hd5'
test_name = f'pixel_pendulum_n_256_steps_100_dt_{dt:.2f}_angle_30-170_vel_-2.00-2.00_len_1.45-1.50_g_12.50-13.00.hd5'
train_path=os.path.join(dataset_dir, train_name)
val_path=os.path.join(dataset_dir, val_name)
test_path=os.path.join(dataset_dir, test_name)
self.datasets['train'] = PixelPendulumDataset(train_path, self.hparams.noise_std)
self.datasets['val'] = PixelPendulumDataset(val_path, self.hparams.noise_std)
self.datasets['test'] = [self.datasets['val'],
PixelPendulumDataset(test_path, self.hparams.noise_std)]
else:
raise Exception(f'Wrong dataset: {self.hparams.dataset}')
print('Train Dataset length: ', len(self.datasets['train']))
print('Val Dataset length: ', len(self.datasets['val']))
print('Test Datasets length: ', [len(dset) for dset in self.datasets['test']])
def train_dataloader(self):
return DataLoader(self.datasets['train'], batch_size=self.hparams.batch_size,
num_workers=self.hparams.num_workers, shuffle=True)
def val_dataloader(self):
return DataLoader(self.datasets['val'], batch_size=self.hparams.batch_size_val,
num_workers=self.hparams.num_workers, shuffle=False)
def test_dataloader(self):
test_dloaders = []
for dset in self.datasets['test']:
test_dloaders.append(DataLoader(dset, batch_size=self.hparams.batch_size_val,
num_workers=self.hparams.num_workers, shuffle=False))
return test_dloaders
def setup(self, stage):
# prepare a train and validation batches
self.batch_sample = {}
self.batch_sample['train'] = next(iter(self.train_dataloader()))
self.batch_sample['val'] = next(iter(self.val_dataloader()))
if self.hparams.dataset=='pendulum_var_length':
self.batch_sample['test_short'] = next(iter(self.test_dataloader()[1]))
self.batch_sample['test_long'] = next(iter(self.test_dataloader()[2]))
if self.hparams.dataset=='pendulum-2':
self.batch_sample['test_out_1'] = next(iter(self.test_dataloader()[3]))
self.batch_sample['test_out_2'] = next(iter(self.test_dataloader()[4]))
elif self.hparams.dataset=='3body-2':
self.batch_sample['test_easy'] = next(iter(self.test_dataloader()[1]))
self.batch_sample['test_hard'] = next(iter(self.test_dataloader()[2]))
else:
# just get a sample from the first test dataset
# index is 1 because 0 is the val dataset
self.batch_sample['test'] = next(iter(self.test_dataloader()[1]))
self.labels_min, self.labels_max = self.datasets['train'].get_labels_min_max()
def configure_optimizers(self):
optimizer = torch.optim.Adam(self.parameters(), lr=self.hparams.learning_rate,
weight_decay=self.hparams.weight_decay)
# try CosineAnnealingLR
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=self.hparams.scheduler_factor,
patience=self.hparams.scheduler_patience,
verbose=True,
threshold=self.hparams.scheduler_threshold,
min_lr=self.hparams.scheduler_min_lr)
return {'optimizer': optimizer,
'lr_scheduler': scheduler,
'monitor': self.hparams.monitor}
def log_histogram(self, name, params):
if self.hparams.use_wandb:
self.logger.experiment.log({name: wandb.Histogram(params.detach())})
else:
self.logger.experiment.add_histogram(name, params, step=self.global_step)
def log_image(self, name, plt_image):
if self.hparams.use_wandb:
self.logger.experiment.log({name: wandb.Image(plt_image)})
else:
pass
# self.logger.experiment.add_image(name, image, self.global_step)
def log_rec_losses(self, batch, stage, rec_loss_sizes, on_epoch=True, on_step=False):
# reconstruction losses for longer trajectories.
max_rollout = np.max(rec_loss_sizes)
start = self.get_start(batch, rec_loss_sizes)
output, target = self.rollout(batch, start=start, rollout_size=max_rollout)
output = output.to(self.device)
target = target.to(self.device)
for step in rec_loss_sizes:
rec_loss = self.reconstruction_loss(output[:, :step], target[:, :step])
self.log(f'{stage}/rec/cumm/{step:04d}', rec_loss, on_step=on_step, on_epoch=on_epoch)
rec_loss = self.reconstruction_loss(output[:, (step-1):step], target[:, (step-1):step])
self.log(f'{stage}/rec/{step:04d}', rec_loss, on_step=on_step, on_epoch=on_epoch)
def _on_after_backward(self): # remove preceding underscore to enable
# used to log parameters grads. pl/wandb has gradient norm logging which is easier/lighter
if self.hparams.debug and ((self.global_step % self.hparams.log_freq) == 0):
for name, params in self.named_parameters():
self.log_histogram(name, params)
self.log_histogram(f'grads/{name}', params.grad.data)
def get_start(self, batch, rec_loss_sizes):
if self.hparams.use_random_start==True:
length = batch['trajectory'].size(1)
max_rollout = np.max(rec_loss_sizes)
max_start = length - self.hparams.model_input_size - max_rollout
start = np.random.choice(range(max_start))
else:
start = 0
return start
def _compute_label_loss(self, labels, mu):
# loss for labels
labels_min, labels_max = self.labels_min, self.labels_max
if self.hparams.sup_loss_type == 'sigmoid':
pred_scaled = torch.sigmoid(mu[:, :self.num_factors]) \
* (labels_max - labels_min) + labels_min
label_loss = F.l1_loss(pred_scaled, labels)
elif self.hparams.sup_loss_type == 'sigmoid_parametrized':
pred_scaled = self.w1 * torch.sigmoid(self.w2 * mu[:, :self.num_factors]) * \
(labels_max - labels_min) + labels_min
label_loss = F.l1_loss(pred_scaled, labels)
self.log('train/sup/w2', self.w2, prog_bar=False, on_step=False, on_epoch=True)
self.log('train/sup/w1', self.w1, prog_bar=False, on_step=False, on_epoch=True)
elif self.hparams.sup_loss_type == 'linear':
pred_scaled = mu[:, :self.num_factors]
label_loss = F.l1_loss(pred_scaled, labels)
elif self.hparams.sup_loss_type == 'linear_scaled':
pred_scaled = mu[:, :self.num_factors] * \
(labels_max - labels_min) + labels_min
label_loss = F.l1_loss(pred_scaled, labels)
elif self.hparams.sup_loss_type == 'BCE':
labels_norm = (labels-labels_min)/(labels_max - labels_min + 1e-6)
BCE = torch.nn.BCEWithLogitsLoss(reduction='mean')
label_loss = BCE(mu[:, :self.num_factors], labels_norm)
else:
raise Warning('Wrong supervised loss type: ', self.hparams.sup_loss_type)
return label_loss